Network Slicing en redes 5G para tráfico Streaming aplicando Deep Learning

Abstract

Network Slicing has been positioned as a technology that is a fundamental pillar for 5G networks, intended to support new technologies with improved requirements in terms of performance and flexibility when dividing the physical network into numerous virtual networks. In this work, an efficient Network Slicing design is proposed using a Deep Neural Network that allows selecting the ideal network Slice so that the packets of each vertical are transported. A training method for the Deep Neural Network is proposed with a metaheuristic algorithm called Glowworm Swarm Optimization (GSO) that presented an efficiency of 99.8% in the accuracy of the data obtained. The virtual network infrastructure is mounted on Openstack – Tacker, which has a complete implementation of what the ETSI suggests for implementations with Virtual Networks and Network Slicing. Free5GC and EURanSim are used to simulate a 5G network environment in the Core of the network and TorchServe is used to deploy the Deep Leaning model in a high-efficiency, low-latency productive environment. The tests focused on comparing the high-resolution Video Streaming vertical, but together with them, Massive IoT traffic and internet browsing were generated to corroborate the correct separation of Slices and the correct flow of traffic in the Slice corresponding to the vertical. An improvement was found in the tests that were done separating the Network Slices and with the Neural Network training method the possibility opens up to new methods that can increasingly provide better accuracy and efficiency when training models even more complex and can help improve the way in which the challenges of implementing networks that support the great diversity of emerging services that are currently being presented are being met. For the Slice deployed for Video Streaming, an improvement in the average transmission speed of 10.71 Mbps is evident.